Watermark detection plays a crucial role in digital watermarking. It has traditionally been tackled using correlation-based techniques. However, correlation-based detection is not the optimum choice either when the host media doesn't follow a Gaussian distribution or when the watermark is not embedded in the host media in an additive way. This paper addresses the problem of DCT (discrete cosine transform) domain multiplicative watermark detection for digital images. First, generalized Gaussian distributions are applied to statistically model the AC (alternative current) DCT coefficients of the original image. Then, the imperceptibility constraint of watermarking is exploited, and watermark detection is formulated as the problem of weak signal detection in non-Gaussian noise. A binary hypothesis test concerning whether or not an image is watermarked is established, and an optimum detection structure for blind watermark detection is derived. Experimental results indicate the superiority of the new detector in the case that the embedding strengths are unknown to the detector. Therefore, the proposed detector can be used for the copyright protection of the digital multimedia data.